Lodestar contracts teams to build AI safety infrastructure that the market will not fund on its own. Specifications and milestones are fixed before payment.
IThe problem
The work that matters most has no buyer. Evaluations that would expose a serious failure mode, or monitoring systems that would catch a dangerous deployment, benefit everyone except the team that builds them. The infrastructure sits in the gap between institutions, and almost none of it gets built.
IICurrent tools
Each was set up for a different job.
IIIMechanism
Lodestar procures safety infrastructure. Each engagement starts from a specification: what the work is and how completion will be judged. Payment follows milestones tied to that specification.
Define the work in enough detail to hold a team to it.
Solicit teams with the operational record to deliver.
Tie payment to milestones, with acceptance criteria written in advance.
Measure success by whether the work is built and used.
IVPrecedent
DARPA was built on the observation that some technologies the public urgently needs will not be produced by the private market on its own. It procures them anyway, through program managers writing structured contracts and paying against milestones.
AI safety has a version of this problem. Important technical work benefits the field broadly while having no commercial buyer to fund it. Lodestar applies the same instrument to it.
VNow
We are talking with the people building, funding, and evaluating this work, and with the institutions that depend on it, to learn where a contracting mechanism would matter most.
If this work intersects yours, please write.